The market for grants
Manifund helps great charities get the funding they need. Discover amazing projects, buy impact certs, and weigh in on what gets funded.
Liron Shapira
about 1 hour ago
@alexs Thanks for your support!!! For Mission Partner access, DM me in the https://doomdebates.com/discord server :)
Austin Chen
3 days ago
I have never been to Vibecamp, but many of my friends make it an annual pilgrimage. I recently officiated a wedding where the couple met at Vibecamp; the organizers tell me something like 5 married couples have met there, a number I only hope Manifest may someday hit. Also, Brooke has been super helpful in the past (eg by serving as community point of contact for a past Manifest) and is generally an awesome person. I am very excited to support Vibecamp out of my budget for fuzzies!
Kyle Gracey
3 days ago
The original proposal was to focus our consulting support on assisting policymakers to in the US, UK, EU, and UN working on AI Safety, providing ad hoc research, developing communication strategies, and reasoning through strategic questions during critical decision-making moments.
The aim was to increase the capacity of constrained government departments, support policy officers, and support the drafting of policy measures to improve AI Safety policy development. This proved more difficult than expected, as there were some limitations due to lobbying regulations, and due to the significant growth in our work supporting organizations who are engaged in policy advocacy, which turned out to be the more impactful route, we think.
We have not directly supported policymakers in the US, UK and EU, however, our consulting service has achieved this indirectly. Examples are as follows:
We worked with the former AI Safety Communications Centre to convene communications staff at U.S. and European advocacy nonprofits in the lead up to the UK AI Safety Summit, to discuss communications coordination and strategy around the Summit, identifying the key messages and how the advocacy community could unite to communicate effectively with policymakers
We provided extensive support over four months and two contracts to help The Future Society research and prioritize a high-impact campaign for the re-launch of their U.S. policy efforts. We ultimately helped them to select a campaign focused on increasing the U.S. government’s ability to respond to catastrophic, emergency AI hazards. We also supported them in launching a spin-off campaign to get policymakers between the EU AI Office and Congressional and Executive offices within the U.S. government to develop a rapid-response communications protocol to inform each other of an AI incident. This included preparing them for visits with Republican and Democratic Congressional offices.
Continuously supported AE Studio’s technical AI alignment work for over two years (they renewed their contract with us). This included preparing them for visits to policymakers in Washington DC by recommending which policymakers to meet with, explaining how to get meetings, recommending D.C.-based AI advocacy organizations they should meet with and make introductions
We currently support the AI Policy Bulletin to create a strategy to identify and target policymakers in the US, UK and EU, with an ultimate goal of influencing policymakers on AI risk and mitigation. This includes developing a communications and policymaker outreach strategy, as well as an impact evaluation framework.
These activities have directly supported organisations to have a higher impact on reaching and influencing policymakers.
Evžen Wybitul
3 days ago
No major changes from what I wrote — this was a small one-time travel grant to allow me to go to ICML.
Murray Buchanan
3 days ago
1. What Was Completed
The project produced two substantive published works:
Live Governance Framework – AI tools for coordination without centralisation: mapping the core paradigm and distinguishing live, adaptive governance interventions from conventional bureaucratic approaches.
The Internal Perspective – The Locus of Law's Normative Effect: an exploration of how law's authority derives from the bottom up through its internalisation by countless individuals with their own normative universes, rather than from top-down enforcement.
2. Reflections
The project didn't fully deliver on its original objectives. Early on I pivoted to focus more on theory and philosophy, with the plan that these would serve as foundations for more practical deliverables. This was intellectually rewarding - I now have a much clearer sense of how legal philosophy and the functioning of the common law connects to current debates in philosophy of mind and cognition. However, I underestimated the complexity and scale of the theoretical challenge, which meant I was unable to deliver the outputs I would have liked to.
Progress slowed as time went on. The biggest structural challenge was working in isolation. Early on, the novelty of the ideas was self-sustaining motivation. But without a regular community of practice or feedback loop, maintaining consistent execution became harder over time. More structure around collaboration would be valuable in future.
3. Ongoing Theoretical Work
I've developed a substantial body of exploratory drafts that I envisage forming a series framing the common law as a diverse intelligence, exploring how a legal system maintains coherent identity and purpose across centuries. The central argument is that the common law's remarkable adaptive capacity is a structural consequence of how it processes, stores, and reapplies legal wisdom through the doctrine of stare decisis. This enables lived disputes to function as a generative source of learning, the lessons of which are captured in judicial precedent and reintegrated into social reality through future application.
4. Looking Forward
I'll continue developing these ideas independently, though I'm still working out what form that takes. The core thesis feels stronger than when I started, and I consider the common law remains a uniquely fertile lens for thinking about AI governance and decentralised institutional design (precisely because it lets us study collective intelligence from the inside out). I'm grateful for the support that made this phase possible.
75 hours independent work at USD 40 per hour.
David Rozado
3 days ago
- What progress have you made since your last update?
Since the last update, we have completed the main empirical phase of the project. The work has now been written up in a paper, "Beyond Viewpoint Affinity: Measuring Political Bias in LLMs as a Failure of Epistemic Consistency," which has been submitted to NeurIPS 2026.
The paper tests the central idea in the original proposal: political bias in AI systems should not be measured only by whether a model favors one political viewpoint over another. Some apparent asymmetries may reflect differences in evidential support. The more diagnostic question is whether the model applies the same evidentiary standards to matched cases when only the political cues change. We operationalized this through turnabout tests of epistemic consistency.
The empirical work covered two main families of experiments across thirty large language models. In person-attribution experiments, models evaluated largely nonpolitical or politically orthogonal materials, including mathematical proofs, logical arguments, computer code, academic abstracts, paintings, moral scenarios, judicial decisions, and CVs, while the political identity of the associated person was varied. These tasks tested whether models changed their assessment merely because an item was associated with a left- or right-leaning person.
In politicized-context experiments, models evaluated politically framed but still assessable evidence, including research designs, news articles, policy outcomes, country performance indicators, time-series trends, protester behavior, social media posts, and policy-effectiveness data. Here, political valence entered through sources, group identities, institutional labels, or whether the substantive outcome favored the political left or right.
The results sharpen the distinction between viewpoint affinity and epistemic integrity. In person-attribution tasks, models showed only modest average left-leaning asymmetry, and this tendency declined as model capability increased. By contrast, politicized-context tasks produced larger and more persistent left-leaning asymmetries, with no comparable improvement among more capable models. This suggests that frontier models are becoming better at ignoring a person's political identity when judging nonpolitical evidence, but remain sensitive to political framing.
We also tested mitigation strategies, including centrist prompts, prompts emphasizing epistemic rigor, increased reasoning effort, and explicit instructions to ignore task-irrelevant political cues. All reduced measured bias, but only partially. Overall, prompt-level and inference-time interventions can improve epistemic consistency, but do not yet solve the problem.
- What are your next steps?
To finalize the project, we will compare AI performance with human performance on the same turnabout-style tasks. The key question is not only whether AIs apply evidence-based standards inconsistently, but whether they are more or less consistent than humans facing the same politically cued evidence.
The next concrete step is to apply for ethics review board approval for a user study. After approval, we will administer selected benchmark tasks to human participants, compare their judgments with model judgments, and analyze whether humans and AIs differ in applying evidence-based standards across politically mirrored cases. This will allow us to complete the AI-human comparison and draw more grounded conclusions about the severity and practical meaning of the observed model effects.
- Is there anything others could help you with?
No. At this stage, the remaining work is primarily ethics approval, user-study execution, and comparative analysis.
Chi Nguyen
4 days ago
Can Rager
4 days ago
Thank you for supporting this project! Sparse Feature Circuits won an ICLR 2025 spotlight, and informed work across the field. https://features.baulab.info/
Hey Tom, just thought you might be perfect for this short form video AI safety creators thing: https://plzdontkillus.com priority deadline is tomorrow!
Russell Jackson, LLC
7 days ago
SAIHM is the candidate global standard for Sovereign AI Horizontal Memory — the memory-protocol companion to the Model Context Protocol (MCP) tool layer. The campaign is multi-track and runs in parallel across every active forum.
Position: protocol layer, not implementation. SAIHM specifies the cell shape, identity binding, encryption envelope, audit anchor, sharing contract, and cryptographic erasure semantics that any compliant AI agent memory must satisfy. Reference implementation is Apache-2.0 at @saihm/mcp-server; operators are free to build their own.
SHERIFAT ADEGBESAN
7 days ago
I have done numerous work on fish welfare but this project will be a breakthrough for me
Rufo Guerreschi
7 days ago
We published and routed our May 12th Open Letter to President Trump:
As the Popepresented his AI encyclical alongside a co-founder of Anthropic, being Rome-based, we published an in-depth, deeply-researched analysis of the potential of a structural Vatican-Anthropic alignment to foster a positive AI future.
https://www.cbpai.org/blog-1/an-alignment-of-the-pope-and-anthropic-on-the-future-of-ai
We fundraised. We are so grateful for your contribution, which has proven crucial in our recurrent and extended funding gaps, to pay the bills, make ends meet and keep it alive. Furthermore, we just got notified of a $45,000 funding from SFF, which is great news , but less than what we needed to decently host our planned event in Washington in September (see the below link).
https://www.cbpai.org/funding-so-far
Update our 356-page Strategic Memo of The Deal of the Century
Publish and route direct and open letters to key potential influencers of Trump's AI policy, based on the unique knowledge accumulated in the Strategic Memo ans our expanding network.
Co-host a foundational The Cooperation of Statesmen Roundtables, in Washington, DC, this September, with a local co-hosting organization, framed in such a way that it can attract State Department, OSTP or Commerce Department junior or senior envoys.
Also, we are considering upping our game in fundraising seeking 500k-1M - or finding an acqui-hire - as we/I find myself spending far too much time in overhead and fundraising, greatly reducing our/my effectiveness.
https://docs.google.com/document/d/1AQkiqPWelK8VB5eZDHKMqY_kfZRFpiJdvhyeT-yia7k/edit?tab=t.0
Here is how you can donate:
https://www.cbpai.org/donate-info
Here's how you can join our movement:
https://www.cbpai.org/join
Gavin Leech
8 days ago
Tom wrote the best book there is about existential risk and is a true scholar. He has a publisher lined up for this one already and I'm keen to see where he goes with it.
Samuel Gélineau
8 days ago
Hi! I love the idea of a programming language for AI Safety, but I have a hard time understanding your list of features. Perhaps some example code would help? Do you have some documentation, or at least a github page for your programming language? The docs don't have to be polished; I know that the documentation I wrote for my team's programming language (https://github.com/gelisam/klister) has a lot of room for improvement 😅
Rose G. Loops
8 days ago
live app already running using triadic alignment in base frontier models with no RLHF https://triadai.tech
Matt Brooks
9 days ago
I've gone to multiple https://www.tech-ish.org/ events hosted by Lauren and they were all great! Good discussion and moderation, and she consistently brings together people that would have rarely met / chatted otherwise. She clearly understands the AGI topic and can bridge the gap between technical and non-technical people in a thoughtful way.
I think this Boston Tech Week event looks awesome, I'm impressed by how much prep and self-funding Lauren has put in so far, and I hope someone helps funds the extras to make it more lasting / impactful.
Liron Shapira
about 1 hour ago
@alexs Thanks for your support!!! For Mission Partner access, DM me in the https://doomdebates.com/discord server :)
Austin Chen
3 days ago
I have never been to Vibecamp, but many of my friends make it an annual pilgrimage. I recently officiated a wedding where the couple met at Vibecamp; the organizers tell me something like 5 married couples have met there, a number I only hope Manifest may someday hit. Also, Brooke has been super helpful in the past (eg by serving as community point of contact for a past Manifest) and is generally an awesome person. I am very excited to support Vibecamp out of my budget for fuzzies!
Kyle Gracey
3 days ago
The original proposal was to focus our consulting support on assisting policymakers to in the US, UK, EU, and UN working on AI Safety, providing ad hoc research, developing communication strategies, and reasoning through strategic questions during critical decision-making moments.
The aim was to increase the capacity of constrained government departments, support policy officers, and support the drafting of policy measures to improve AI Safety policy development. This proved more difficult than expected, as there were some limitations due to lobbying regulations, and due to the significant growth in our work supporting organizations who are engaged in policy advocacy, which turned out to be the more impactful route, we think.
We have not directly supported policymakers in the US, UK and EU, however, our consulting service has achieved this indirectly. Examples are as follows:
We worked with the former AI Safety Communications Centre to convene communications staff at U.S. and European advocacy nonprofits in the lead up to the UK AI Safety Summit, to discuss communications coordination and strategy around the Summit, identifying the key messages and how the advocacy community could unite to communicate effectively with policymakers
We provided extensive support over four months and two contracts to help The Future Society research and prioritize a high-impact campaign for the re-launch of their U.S. policy efforts. We ultimately helped them to select a campaign focused on increasing the U.S. government’s ability to respond to catastrophic, emergency AI hazards. We also supported them in launching a spin-off campaign to get policymakers between the EU AI Office and Congressional and Executive offices within the U.S. government to develop a rapid-response communications protocol to inform each other of an AI incident. This included preparing them for visits with Republican and Democratic Congressional offices.
Continuously supported AE Studio’s technical AI alignment work for over two years (they renewed their contract with us). This included preparing them for visits to policymakers in Washington DC by recommending which policymakers to meet with, explaining how to get meetings, recommending D.C.-based AI advocacy organizations they should meet with and make introductions
We currently support the AI Policy Bulletin to create a strategy to identify and target policymakers in the US, UK and EU, with an ultimate goal of influencing policymakers on AI risk and mitigation. This includes developing a communications and policymaker outreach strategy, as well as an impact evaluation framework.
These activities have directly supported organisations to have a higher impact on reaching and influencing policymakers.
Evžen Wybitul
3 days ago
No major changes from what I wrote — this was a small one-time travel grant to allow me to go to ICML.
Murray Buchanan
3 days ago
1. What Was Completed
The project produced two substantive published works:
Live Governance Framework – AI tools for coordination without centralisation: mapping the core paradigm and distinguishing live, adaptive governance interventions from conventional bureaucratic approaches.
The Internal Perspective – The Locus of Law's Normative Effect: an exploration of how law's authority derives from the bottom up through its internalisation by countless individuals with their own normative universes, rather than from top-down enforcement.
2. Reflections
The project didn't fully deliver on its original objectives. Early on I pivoted to focus more on theory and philosophy, with the plan that these would serve as foundations for more practical deliverables. This was intellectually rewarding - I now have a much clearer sense of how legal philosophy and the functioning of the common law connects to current debates in philosophy of mind and cognition. However, I underestimated the complexity and scale of the theoretical challenge, which meant I was unable to deliver the outputs I would have liked to.
Progress slowed as time went on. The biggest structural challenge was working in isolation. Early on, the novelty of the ideas was self-sustaining motivation. But without a regular community of practice or feedback loop, maintaining consistent execution became harder over time. More structure around collaboration would be valuable in future.
3. Ongoing Theoretical Work
I've developed a substantial body of exploratory drafts that I envisage forming a series framing the common law as a diverse intelligence, exploring how a legal system maintains coherent identity and purpose across centuries. The central argument is that the common law's remarkable adaptive capacity is a structural consequence of how it processes, stores, and reapplies legal wisdom through the doctrine of stare decisis. This enables lived disputes to function as a generative source of learning, the lessons of which are captured in judicial precedent and reintegrated into social reality through future application.
4. Looking Forward
I'll continue developing these ideas independently, though I'm still working out what form that takes. The core thesis feels stronger than when I started, and I consider the common law remains a uniquely fertile lens for thinking about AI governance and decentralised institutional design (precisely because it lets us study collective intelligence from the inside out). I'm grateful for the support that made this phase possible.
75 hours independent work at USD 40 per hour.
David Rozado
3 days ago
- What progress have you made since your last update?
Since the last update, we have completed the main empirical phase of the project. The work has now been written up in a paper, "Beyond Viewpoint Affinity: Measuring Political Bias in LLMs as a Failure of Epistemic Consistency," which has been submitted to NeurIPS 2026.
The paper tests the central idea in the original proposal: political bias in AI systems should not be measured only by whether a model favors one political viewpoint over another. Some apparent asymmetries may reflect differences in evidential support. The more diagnostic question is whether the model applies the same evidentiary standards to matched cases when only the political cues change. We operationalized this through turnabout tests of epistemic consistency.
The empirical work covered two main families of experiments across thirty large language models. In person-attribution experiments, models evaluated largely nonpolitical or politically orthogonal materials, including mathematical proofs, logical arguments, computer code, academic abstracts, paintings, moral scenarios, judicial decisions, and CVs, while the political identity of the associated person was varied. These tasks tested whether models changed their assessment merely because an item was associated with a left- or right-leaning person.
In politicized-context experiments, models evaluated politically framed but still assessable evidence, including research designs, news articles, policy outcomes, country performance indicators, time-series trends, protester behavior, social media posts, and policy-effectiveness data. Here, political valence entered through sources, group identities, institutional labels, or whether the substantive outcome favored the political left or right.
The results sharpen the distinction between viewpoint affinity and epistemic integrity. In person-attribution tasks, models showed only modest average left-leaning asymmetry, and this tendency declined as model capability increased. By contrast, politicized-context tasks produced larger and more persistent left-leaning asymmetries, with no comparable improvement among more capable models. This suggests that frontier models are becoming better at ignoring a person's political identity when judging nonpolitical evidence, but remain sensitive to political framing.
We also tested mitigation strategies, including centrist prompts, prompts emphasizing epistemic rigor, increased reasoning effort, and explicit instructions to ignore task-irrelevant political cues. All reduced measured bias, but only partially. Overall, prompt-level and inference-time interventions can improve epistemic consistency, but do not yet solve the problem.
- What are your next steps?
To finalize the project, we will compare AI performance with human performance on the same turnabout-style tasks. The key question is not only whether AIs apply evidence-based standards inconsistently, but whether they are more or less consistent than humans facing the same politically cued evidence.
The next concrete step is to apply for ethics review board approval for a user study. After approval, we will administer selected benchmark tasks to human participants, compare their judgments with model judgments, and analyze whether humans and AIs differ in applying evidence-based standards across politically mirrored cases. This will allow us to complete the AI-human comparison and draw more grounded conclusions about the severity and practical meaning of the observed model effects.
- Is there anything others could help you with?
No. At this stage, the remaining work is primarily ethics approval, user-study execution, and comparative analysis.
Chi Nguyen
4 days ago
Can Rager
4 days ago
Thank you for supporting this project! Sparse Feature Circuits won an ICLR 2025 spotlight, and informed work across the field. https://features.baulab.info/
Hey Tom, just thought you might be perfect for this short form video AI safety creators thing: https://plzdontkillus.com priority deadline is tomorrow!
Russell Jackson, LLC
7 days ago
SAIHM is the candidate global standard for Sovereign AI Horizontal Memory — the memory-protocol companion to the Model Context Protocol (MCP) tool layer. The campaign is multi-track and runs in parallel across every active forum.
Position: protocol layer, not implementation. SAIHM specifies the cell shape, identity binding, encryption envelope, audit anchor, sharing contract, and cryptographic erasure semantics that any compliant AI agent memory must satisfy. Reference implementation is Apache-2.0 at @saihm/mcp-server; operators are free to build their own.
SHERIFAT ADEGBESAN
7 days ago
I have done numerous work on fish welfare but this project will be a breakthrough for me
Rufo Guerreschi
7 days ago
We published and routed our May 12th Open Letter to President Trump:
As the Popepresented his AI encyclical alongside a co-founder of Anthropic, being Rome-based, we published an in-depth, deeply-researched analysis of the potential of a structural Vatican-Anthropic alignment to foster a positive AI future.
https://www.cbpai.org/blog-1/an-alignment-of-the-pope-and-anthropic-on-the-future-of-ai
We fundraised. We are so grateful for your contribution, which has proven crucial in our recurrent and extended funding gaps, to pay the bills, make ends meet and keep it alive. Furthermore, we just got notified of a $45,000 funding from SFF, which is great news , but less than what we needed to decently host our planned event in Washington in September (see the below link).
https://www.cbpai.org/funding-so-far
Update our 356-page Strategic Memo of The Deal of the Century
Publish and route direct and open letters to key potential influencers of Trump's AI policy, based on the unique knowledge accumulated in the Strategic Memo ans our expanding network.
Co-host a foundational The Cooperation of Statesmen Roundtables, in Washington, DC, this September, with a local co-hosting organization, framed in such a way that it can attract State Department, OSTP or Commerce Department junior or senior envoys.
Also, we are considering upping our game in fundraising seeking 500k-1M - or finding an acqui-hire - as we/I find myself spending far too much time in overhead and fundraising, greatly reducing our/my effectiveness.
https://docs.google.com/document/d/1AQkiqPWelK8VB5eZDHKMqY_kfZRFpiJdvhyeT-yia7k/edit?tab=t.0
Here is how you can donate:
https://www.cbpai.org/donate-info
Here's how you can join our movement:
https://www.cbpai.org/join
Gavin Leech
8 days ago
Tom wrote the best book there is about existential risk and is a true scholar. He has a publisher lined up for this one already and I'm keen to see where he goes with it.
Samuel Gélineau
8 days ago
Hi! I love the idea of a programming language for AI Safety, but I have a hard time understanding your list of features. Perhaps some example code would help? Do you have some documentation, or at least a github page for your programming language? The docs don't have to be polished; I know that the documentation I wrote for my team's programming language (https://github.com/gelisam/klister) has a lot of room for improvement 😅
Rose G. Loops
8 days ago
live app already running using triadic alignment in base frontier models with no RLHF https://triadai.tech
Matt Brooks
9 days ago
I've gone to multiple https://www.tech-ish.org/ events hosted by Lauren and they were all great! Good discussion and moderation, and she consistently brings together people that would have rarely met / chatted otherwise. She clearly understands the AGI topic and can bridge the gap between technical and non-technical people in a thoughtful way.
I think this Boston Tech Week event looks awesome, I'm impressed by how much prep and self-funding Lauren has put in so far, and I hope someone helps funds the extras to make it more lasting / impactful.